A pointer meter reading method based on human-like reading sequence and keypoint detection

Abstract

To aid in the development of unmanned factories and increase industrial production efficiency, meter recognition reading methods based on machine vision are replacing manual meter reading. This paper proposes a recognition and reading method for pointer-type meters based on the lightweight networks YOLOv8S and MC-DeeplabV3Plus, with the goal of addressing existing methods’ poor robustness and low reading accuracy on edge devices. It applies to pointer-type meters with uniformly distributed scales. The proposed Channel Depth-wise Convolutional Attention (CDCA) module improves the channel attention module’s accuracy in segmenting details and edge features. It is integrated into the DeeplabV3Plus network alongside the Mixed Local Channel Attention (MLCA) module, thereby improving the model’s segmentation performance in complex scenarios. At the same time, MobileNetV2 is selected as the segmentation network’s backbone due to its lightweight structure, which makes it suitable for deployment on devices with limited resources. To enhance the stability of meter readings, this paper uses a flexible angular approach to calculate the readings. This method acquires the meter’s key points by mimicking the human reading sequence and maintains good robustness even when partial information is missing. The experimental results demonstrate that this method achieves a fiducial error of approximately 0.039 % in an interference-free laboratory environment and 0.733 % in real-world scenarios, and that the average frame rate for single image processing without GPU support is 3.61 FPS with only 14.18 million parameters, indicating a high application potential. The code is available at: https://github.com/paopao6777/det-read-pointer-meter.